Kimi K2.6
Moonshot AI's open-weight 1-trillion-parameter coding model — it ties GPT-5.5 on SWE-bench Pro while costing roughly 80% less.
Kimi K2.6 specs at a glance
| Maker | Moonshot AI |
| Released | 20 April 2026 |
| Architecture | MoE · 1T total params · 32B active · 384 experts (8 + 1 shared) · 61 layers |
| Attention | Multi-head Latent Attention · SwiGLU activation |
| Context window | 262,144 tokens (max output 262,142) |
| Modalities | Text, images, video — one architecture |
| API pricing | $0.60 / $2.50 per 1M input / output tokens (official) |
| Licence | Modified MIT — open weights |
| Best at | Long-horizon coding, multi-agent orchestration |
What Kimi K2.6 is built for
Kimi K2.6 is a coding-first model. Moonshot AI designed it for long-horizon coding — complex, end-to-end tasks that span many steps — along with coding-driven UI/UX generation and multi-agent orchestration, handling work across Python, Rust and Go. It was the most-read model release of April 2026, and the reason is simple: it matches a Western flagship on the hardest engineering benchmark while being open-weight and far cheaper.
Under the hood it is a large Mixture-of-Experts model: 1 trillion total parameters but only 32B active per token, drawn from 384 experts (8 selected plus 1 shared) across 61 transformer layers. It also processes text, images and video in the same architecture, without bolt-on vision modules.
Kimi K2.6 benchmarks
| Benchmark | Kimi K2.6 | Notes |
|---|---|---|
| SWE-bench Pro | 58.6% | Ties GPT-5.5; beats GPT-5.4 (57.7%) and Opus 4.6 (53.4%) |
| Humanity's Last Exam (with tools) | 54.0% | Leads every frontier model Moonshot compared against |
Key takeaway
Kimi K2.6 matching GPT-5.5's 58.6% on SWE-bench Pro — at roughly 80% lower cost and with open weights — is the headline result. For coding-heavy budgets, it reshapes the price-performance calculation.
Pricing
On Moonshot's official API, Kimi K2.6 costs $0.60 per 1M input tokens and $2.50 per 1M output tokens. Third-party providers price it a little higher — OpenRouter lists $0.73 / $3.49 — and across the nine tracked providers blended pricing runs roughly $1.15 to $2.15 per 1M tokens. Either way it sits far below GPT-5.5's $5 / $30, which is where the "saves ~80% on coding costs" claim comes from.
Licence and Kimi Code
Kimi K2.6 ships with open weights under a modified MIT licence. In practice it behaves as standard MIT — no royalties, free to modify and self-host — with one clause: if you deploy K2.6 in a commercial product exceeding 100 million monthly active users or $20 million in monthly revenue, you must prominently display "Kimi K2" in the user interface. Below those thresholds the attribution requirement does not apply.
Moonshot also ships Kimi Code, a coding agent and CLI built around the K2.6 model — the intended way to use it for end-to-end, long-horizon engineering tasks.
Who should use Kimi K2.6
- Coding-heavy teams on a budget — frontier SWE-bench scores at ~80% lower cost.
- Developers building agentic coding workflows — long-horizon, multi-agent orchestration is its design target.
- Anyone who wants open weights — self-hostable, modified-MIT, no lock-in for the vast majority of users.
How Kimi K2.6 compares
- GPT-5.5 ("Spud") — the model it ties on SWE-bench Pro →
- DeepSeek V4 — the other major open coding model →
- Gemini 3.1 Ultra overview →
Frequently asked questions
When was Kimi K2.6 released?
Moonshot AI released Kimi K2.6 on 20 April 2026. It is an open-weight 1-trillion-parameter model built for long-horizon coding and multi-agent work.
How much does Kimi K2.6 cost?
On Moonshot's official API, Kimi K2.6 costs $0.60 per million input tokens and $2.50 per million output tokens. Third-party providers such as OpenRouter price it slightly higher, around $0.73 / $3.49. It costs roughly 80% less than GPT-5.5 for comparable coding performance.
Is Kimi K2.6 open source?
Kimi K2.6 has open weights under a modified MIT licence. It functions as standard MIT unless you deploy it in a commercial product with over 100 million monthly active users or more than $20 million in monthly revenue, in which case you must display "Kimi K2" in the user interface.
What is Kimi K2.6's context window?
Kimi K2.6 has a 262,144 token context window, with a maximum output of 262,142 tokens.
How good is Kimi K2.6 at coding?
Kimi K2.6 scores 58.6% on SWE-bench Pro — tying GPT-5.5 and beating GPT-5.4 and Claude Opus 4.6. It also leads on Humanity's Last Exam with tools at 54.0%.
What is Kimi Code?
Kimi Code is Moonshot AI's coding agent and CLI built around the Kimi K2.6 model, aimed at long-horizon, end-to-end coding tasks across languages including Python, Rust and Go.